Tamir Dresher: Squad Agent Workflows - Episode 407

Speaker A: Squad Agent workflows with Tamir Drescher.

Speaker B: The AI DevOps podcast is a show for those shipping software using AI Net, Azure and DevOps. Each show brings you hard hitting interviews with industry experts, innovating better methods and sharing success stories. Sponsoring the podcast is ClearMeasure, a software architecture and engineering firm that implements AI to empower software teams to establish quality, achieve stability and increase speed. And now your host, Jeffrey Palermo.

Speaker A: Welcome to the show. I'm Jeffrey Palermo, your host for helping you and your teams move fast and deliver quality and to run your software with confidence in Azure, all while using everything that AI, Azure and the. NET ecosystem has to offer. My guest today on the show is Tamir Drescher and he is a principal engineer at Microsoft Threat Protection where he focuses on scaling AI agent systems and distributed architectures. He brings over 15 years of experience building large scale distributed systems and he's the co creator of an open source multi agent runtime for GitHub Copilot that uh, orchestrates AI teams directly inside your repository. Tamir uh, is also the author of RX. Net in Action, a book from Manning and also hands on Full Stack web development with ASP NET Core, a book from Packet and he's been a lecturer in software engineering at the Rupin academic center since 2013. He's also a prominent figure in the Israeli and international developer communities and he's a Microsoft MVP alumnus before joining Microsoft and he speaks frequently at global conferences and writes actively on his blog@tamirdresher.com, we'll include a link to that in the show notes. But Tamir, welcome to the show. How are you sir?

Speaker C: I'm great. Thank you for having me Jeffrey. Uh, it's a pleasure to be here.

Speaker A: Pleasure uh, is all mine. I'm excited uh you work on a team with Brady uh Gaster and he introduced uh us and before we get into uh, get into uh what you're working on, which I think, which I think is amazing, I always like to ask our guests to go kind of back in time just so that our listeners can get to know you a little bit and ask if you remember what was it that originally got you into the field of software engineering and captured your, captured your imagination to go into this field.

Speaker C: So uh, since I was a child I was always surrounded by you know, electronics and games and I don't know I got my uh first uh, PC when I was around five years old. Um and with my um uncles and uh cousins we like did a lot of uh hacking uh stuff and playing games. And uh, back then the movie Hackers, uh, came out, uh, and you know, just kind of make uh, your imagination. And my family was really, really into, uh, Star Trek. So that's uh, another thing that had a lot of influence on me. Um, and, and so it just became natural that this is what I want to do. Um, and um, went into studies, uh, started working on that. Um, and it's kind of funny because I got the very first copy of Visual, um, Studio and Net, uh, from a friend of my father, which was a Microsoft employee from the US and he came to visit and he bought me like a giant book. Really. It was, uh, I don't know, like a thousand pages with a Cindy for Visual Studio. And that's how I got into Net. Uh, I was really young. Um, and it was really fun.

Speaker A: Nice. Nice. And, And I'm guessing it has stayed fun ever since.

Speaker C: Yes, ever since. It's uh, you know, just the thing that keeps me, keeps me happy.

Speaker A: Awesome. Awesome. Well, you've been doing some really interesting work with Squad. And I, I don't think, uh. I think there's a lot of listeners who still have never used Squad in any form for, for at all. And so can I ask you to start with the vision behind Squad overall? What is lacking with Copilot, all the other AI tools? And why is something like Squad needed? What's the vision?

Speaker C: Absolutely. So I think it's good to take uh, like a glimpse at something outside of uh, AI first that, you know, let's say uh, doctors in, uh, operation room. Right. Um, when you work today with a single AI agent is like saying that inside the operation room you will have a single doctor that read everything on, on the planet. Um, and you know, it's um, he has all the knowledge, but it's just a single entity operating. Um, now with something like that, as a doctor, we know that you cannot really know everything. You get more generalist than with speciality. And over time you get this type of fatigue. Um, and as you go, the performance will be lower. And so, uh, the vision of Squad is that let's break that AI into a team of specialists instead of one generalist and have something that knows how to coordinate them and let them learn and evolve over time. And beside that, our vision is to not only have a team of AI, but building this bridge between human and AI working together as a, as a single team collaborating. So some of the work is being done by the team members that are AI specialized in specific parts, and the other parts are being done by the human and all of this just work together in some kind of a symbiotic relationship. Um, and having all the ceremonies and all the procedures and everything needed for this type of team to really work together and evolve and learn. Uh, that's basically our vision and making all that not just for engineering, like democratize AI, allowing everyone on the planet to build their own AI assistants, a team that can serve them and help them produce more. This is what we are trying to achieve. Uh, and we had a lot of uh, success. And when I see someone outside of engineering, outside of uh, programming using Squad for their project, really that's the thing that makes me go with a smile on my face when I go uh, to sleep at night.

Speaker A: Well I'd like to talk to you about that, the non engineering aspect of it because there's a lot of programmers who are talking about creating specialized prompts to create specialized agents that do different aspects of coding. But, but you focused on the non engineer workflow. Uh, so uh, what is, what does that look like? Can you give some examples that are, that you've seen already out there in the wild?

Speaker C: Absolutely. I can share from my own experience uh, and also things that I uh, started to see. So one of the things that um, a few months ago I wanted to see how I can leverage AI is to uh, seek more information about the history of my grandfather which was in the Holocaust, uh, and we had no real traces on part of the family. And so you have a lot of people doing genealogy investigating. But I try to do it myself and I found that with AI, uh you get those type of hallucinations, it gets into those loops and using Squad on top of that and building specialized um team members focusing on specific areas like um, um genealogy, others on specific areas of Europe and stuff like that. It managed in a very short while um, to do this cross sectioning between different databases and went into a lot of places on the web and even contacted people, uh, people and collected all this information and produced real things like new material that I didn't have. Um, and so when I did that I said okay, this is something that can serve others. And I kind of created this type of a squat template uh for Holocaust research, um, and published it open source. Um, and then someone I know used uh, that and they discovered uh, like uh, a lost cousin. Like 75 years they've been searching, uh, and they couldn't find. And this Squad was able to find like an article in the New York Times about the funeral and then connected it and saw that the person changed the name and all of that finding the person uh, they were searching for. So this is one area when someone can do a research and having specialized agent running a parallel doing a lot of work. You can take that example to other pieces. I know about people searching for lost art. You know during worlds art is being stolen and then they pop in different places. So this team of researchers is something else beside that. Um, one of the um, like aspects of that of democratizing AI is how can you you know, expose the AI, ah, in a simple way to people that are not like tech savvy, someone who doesn't really into the details. Uh, and one of those samples is uh, created this squad for my mother, uh, in law. Uh, she has an online store and basically when she wants to change something in the products and everything she contacts me or my wife. And so we created this squad that uh, she communicates over email with the squad, uh, passing like with a normal human. Just these are the products and the images and the squad will go and do the changes. She asked for um, on the online store and changing the website. Uh, at first she didn't realize it was AI, she thought it was me. Um, other people using that for searching uh, for like um, experiments, uh, medical experiments where you as someone who has cancer, uh, or something, they are searching for those type of uh, researchers uh, and it's really hard to know how to really explain it or search for that. That's another area where AI team of AI experts can find and make it easier for someone who don't have the knowledge what is out there and can really uh, sometimes uh, save their life. So those are the type of how we can create AI teams operating and making it easy for someone to walk, uh, sometimes without even realizing. Ah, it's AI.

Speaker A: Okay, so let's get technical for just a minute. Um, if someone wants to use this, how do they get started? What exactly do they download? What do they install, what do they connect?

Speaker C: So Squad is a framework on top of Copilot, which we support. Um, and so in essence the squad itself is a custom agent for Copilot. This uh, custom agent knows all the definitions about how to work in a team, how to delegate role collecting work coordinates, the different sub agents. Uh, and the squad itself is uh, actually uh, you can say a folder inside your repo with a structure where you have all the sub agents and their charters, what they are allowed to do, what is the definition and specialty, their personality, which we'll talk about uh, a bit later. Um, and a Bunch of other folders with all the decisions and the knowledge and the routing rules and stuff like that. And so to get all this structural folders into your repo, the thing you need to do is just install the um, squad cli, which is an NPM package. Um, and then all you need to do is just run the squad, uh, init inside your repo. This will just scaffold all those files into your repo. And from that moment on, if you don't want to use the CLI anymore, um, you can throw it away because basically from that moment on, all the, you know, uh, procedures, everything, all the workflows are basically kind of built inside the custom agent and all the m config files that we put inside your repo. And then you just run copilot. You select the custom agent, basically just doing a slash agent in your copilot session, selecting squad and start to communicate with it the first time you run it. When you kind of the squad introduce itself, it will see if this is an uh, empty repo or something that has already uh, code inside. Um, and then we'll kind of suggest the theme that uh, should be part uh, of your project. If it's uh, um, an existing project, it will find what is the technologies, what is the uh, guidelines, the convention, stuff like that. If it's something new, it will ask you what are you trying to do here? And uh, you can define the project whether it's a game, an app or whatever. Um, and then the squad will do this uh, first uh, part analysis. And we suggest these are your team members, um, each one with a specific um, specialty in technology overall. On top of that we uh, allow the user to select the universe from where those members will be picked. What we found in some other experiments and research we did was that just like normal, uh, human teams, when you have diversity inside the team, it makes the entire team better, much more efficient and getting more quality work. Um, and so you can select a universe like the Matrix movie or uh, Marvel, uh, superheroes or Star Trek or Star wars or whatever you like. And then each of those team members that you have will be based on some character, something, um, and this type of personality that you add on top of this sub agent with the role it needs to perform makes the entire interaction between the entire team work differently and produce better results. Um, which is also fun to see, right, because you see them discuss, they add jokes or whatever needed, uh, to work as a uh, as a team together. So these are like the very minimal set of actions you need to do to start working with a squad.

Speaker A: Okay, so this, on the face of it, it sounds uh, similar to some of the other AI agents as far as just having a prompt that defines the agent's specialty and helps it focus. But I want to ask you about what I think makes Squad very differentiated and that's the um, JavaScript extensions or your blog post. Uh, Ralph gets a tool belt, um, and adding on and we'll include a link to that in the show notes, but adding on deterministic script extensions that each of these squad, uh, agent specialists can use. Uh, because I think that anybody who's tried to use custom agents has realized that if the agent is defined by just a prompt, then architecturally it's non deterministic. And at certain times we just need to insert code in there because we need something to be deterministic to do exactly the same thing every time and not just be based on a prompt inference. So can, can you kind of share what is the, what's the architecture of the extensions and how do those work?

Speaker C: So I think that like most people working with AI today, what they have in mind is that you need to open something like the copilot or um, you know, Claude or something and you know, have an interactive session with it. But what you really want is to have some ability for the AI, whether it's quad or something else, to just run in the background and autonomously start doing things. And when needed you start to interact, give it, you know, more tasks getting uh, reviewed, the artifacts and stuff like that. So um, a while ago, like a long time ago in, you know, in the time that squad is alive, um, we had this concept of uh, uh, as part of the cli, we call it the Squad Watch. And when you run the Squad watch, the squad itself goes into this type of a loop with deterministic steps and some steps that are not deterministic where the squad itself really leverage the LLM and start doing some of the intelligent stuff. And those steps can be like the deterministic steps are. Okay, let's see if there's any walk in the backlog. Let's walk with the GitHub, uh, backlog or the Azure DevOps, uh, uh, walk items and see if there's any walk. Let's filter the walk for those that are for this uh, squad and our um, thinks that we can do it on this specific machine because we support capability routing. Um, and then, okay, let's take those issues and then let's spawn the copilot with the squad to Start working on that. And then let's maybe go and uh, scan the emails, scan uh, teams, scan whatever channels that you have to see if there's any input that needs to be taken into another task and then do some kind of a, ah, bookkeeping, change statuses, update the board. Right. So some of the steps require the LLM, others are not. And it doesn't make sense to first pay for those, you know, tokens for something that is deterministic. Second, you want to have some kind of a governance on top of that. So that was the basic idea of Squad Watch. Uh, later on uh, we got some, you know, um, uh, some requests from the community because not everybody works in the very same way. And on top of your squad you want uh, to add more steps that are deterministic to do certain things, whether it's to go and also query some other system for tickets or doing some kind of a cleanup or you know, running some kind of applications or uh, just you know, um, making an email by a format and send it uh, through the CRM system that you want. And so on top of what I just mentioned, with the Squad Watch, you can basically add those steps. And like the design we selected for that is that as part of your repo, where you have the squad folder, you can put Those uh, um, JavaScript, you know, files with the steps, with the specific contract that we define and the Squad Watch will know how to upload them based on the configuration. We'll stitch that into the workflow that we're already running and then you can have the entire squad also get the output from the previous steps to work on that. And so this is for the, you know, where in the past we would say I won't write a full system application. I can write a script that do something for me or just open an Excel and do some uh, you know, formulas in the spreadsheet. So you can say this is like the equivalent in the AI world. Some of the things I just want the Squad or the AI based on my prompt to do. Others I want to do with real coding, real deterministic steps. But I want the entire thing to work together. And I think part of our vision is not only how we collaborate, you know, the humans and the AI, but also how to make the AI like seamless, just any other utility that you want, uh, in your uh, terminal and pipe uh, the output from a previous uh, to the next command where uh, incidentally this command is doing some intelligence, uh, through some AI and Squad.

Speaker A: Now on the engineering side, one of the real common loops is someone commits something and creates a pull request and submits a pull request. But there's always this little bit of waiting. It could be a short period of time where we don't know if the build is green or if the build broke on the pull request. And so it sounds like, and I think a lot of people are using the command line, uh, tools to say monitor the build on the pull request and make sure that it's green. But then all that console output and it's, it's using up, it's using up AI tokens probably, you know, unnecessarily. Uh, it seems like that would be just one of the low hanging fruit scenarios that these deterministic extensions could handle. Um, am I thinking about that correctly?

Speaker C: Yeah, yeah, absolutely. I mean one of the steps you would want with something like that that runs in the background is to go and pull the latest changes from your repo, right? Like this is an operation, you don't need a prompt to say pull latest, right? You can add something that just do the git pull. So those type of things, doing those really, really simple conditions, simple checks, validations, running those tasks that you know you need to do, they are part of the process. You, you, there's no like judgment uh call here. You need to do that. Those are the perfect things to put as part of the deterministic steps, um, and have it embedded as part of the entire process. That way you save tokens and that way you also make it more efficient. It would be much faster than waiting for the LLM to yield, how to do and then run it and so on. I think that one thing that maybe I didn't stress enough is that the differentiator of squad versus others is that I mentioned the squad folder and the squad folder is not just a static bunch of files where you have the definition of the squad. This is where all the learnings and you know, all those you can say constitutions is, is baked in. And as the squad work and as, as it's getting more inputs from the, the other team members and doing the work on the task themselves, they learn, they have insights and they are making decisions. Or if you are giving them a uh, like no, no longer uh, should you do that. From now on just do this. This is a directive, they will put it inside their files and the squad itself will make sure they are not breaking this type of uh, contract uh, that you have. So it's like hiring a real team and you have this type of a learning phase and over Time they become better until you don't need to educate them anymore. They just do the work as you expect them to do. This is how squad work. The more you work with it, the more it's doing this type of processing, the better it will be. And it's self learning by itself. And that's another thing uh, I think is kind of a differentiator. And yes, everything lives inside your repo. So squad has memory, Squad has a memory. Uh and we are doing a lot of job of work right now to see how we can adjust and fine tune the memory, uh, stuff. Um, because when we say memory it's not just you know, remembering what we did in the past. Right. It's also understanding how we need to do things for the future. And like over time we kind of realized that not every memory piece is the same. Sometimes you have things that you need to remember only for a specific session. Sometimes it's only for a specific walk stream something specific that you work on even if it's over time with pausing and resuming. And some of the things are long running. And so for those different types of memory items you have different type of processing you need to do where you need to retrieve only the relevant memory for a specific operation. And I think another, another like aspect that uh listeners need to understand. When we say sub agents, I mean when the squad is a custom engine work and decide that now this is a uh, trust it needs to delegate to sub agent. The spawning of the sub agent starts with a uh, clean fresh context. All it knows is what is uh, by itself uh the squad member. And so now is the perfect phase to fill it up with the relevant information. And you want to be efficient, you want the thing compressed enough so it wouldn't take too much of the context. But you know, not lossy enough or too much so you wouldn't load. So we added on top of the uh, squad CLI a few of those tools that when the squad is running and needs to retrieve from its information it can use these tools. And again these tools are deterministic, it's by code. It just knows how to filter, digest, retrieve, uh, and filter. And of course like I said before, you don't have to use that if you prefer not to use any CLI tool as part of the only copilot. By itself it will just result to the regular file system but you might get more information that you needed. Also part of those tools is allowing the squad to forget. You don't need to remember things forever. Some things are uh, for A period of time they are only relevant until a certain event, until something happened. So part of the processing that the tools knows how to do is that when there is a time limit for something we will know how to delete it and you know, not load it just because it's there. So those are type of the thing we are working on now to make it more adjustable and again making sure the squad will always be focused and each sub agent focused on the task and nothing else.

Speaker A: Okay, so some of your examples in your blog post on uh, this have nothing to do with programming but are just business workflows, non engineering workflows, even something like uh, monitoring a customer record in a CRM system or monitoring some type of uh, an email inbox. So is your vision for this that we can literally watch anything that has an API of some sort and then just do whatever's necessary? Is it that broadcast?

Speaker C: I think my philosophy is that um, AI agents are uh, like the new microservices and we need to treat them the same way. With everything we learned over time about distributed systems and coordination and you know, synchronization, um, all of that are the same relevant to AI. It's just that now what the service or the component is doing as part of the process has some kind of uh, intelligence and decision making. Um, and it can make those type of a judgment calls. And so when we think about a business process there are many places where you need to make a call. What we used to do as you know, engineer is define this type of business uh, rule engine or something that just knows how to do this type of condition. So now we can have something even stronger. But not only is it stronger because it just knows to look at some specific context and information and get you to the decision. It can really, really learn from the mistakes it did, right? And instead of waiting for the next release where somebody will add a new rule, now it can do it on the fly, right? So to achieve that specific, you know, uh, world um, we need to have the tool sets to make it so one of the things for example that uh, I'm working on and it's almost done, uh, and I did write about it in the blog, is that um, using agent framework, right, where you can basically build those type of durable workflows where some of those are AI agents now you can have an NI agent which is a squad. And so now everything you know about AI agent as part of a process can be with this type of uh, self learning, evolving entity which in like the most naive way all the files are just inside, you know, the service. But you can save. Uh, nobody really said that uh, uh, it has to be in the repo from the starter. We created this type of abstraction where the decision and all the Squad data can also be an external database for engineering. Um, you know it's very easy to have it in the repo, but for a system you can have Squad running but just using uh, an external, you know, persistent layer, um, but still be something that the squad itself will kind of build itself. One of the nice or interesting thing to see is when the squad itself decide that it needs to hire a new team member, there's something new, we didn't anticipate it before and it's better to have a dedicated sub agent focused on that. And more interesting than that, if we run this type of a ceremony every once in a while of uh, retrospective and the squad can decide to fire a member because it is not performing well enough or it didn't do the right things and so it's better to get rid of that sub agent uh, instead of keeping that. So just like teams, they have a retro and they just work together to get the best result.

Speaker A: So. So you mentioned the Agent framework. Are you specifically talking about creating an application that uses the Microsoft Agents AI Foundry package?

Speaker C: Not necessarily Foundry, but yes, using the Microsoft Agent framework, which is the next you can say version of semantic kernel, um, which give you this higher level of abstraction on agents. And yes, every one of those agents essentially will use the Microsoft AI extensions. But you can now build a full workflow and really connect the different steps. Some of the steps are deterministic in your application, some are not. Some can only do this type of indexing and uh, put it in a vector database and others need to retrieve and make a decision. But now um, you can combine everything under a single cover in a. NET application. Um, um, Agent framework, by the way also support Python but basically not looking at the squad again just like a team. It's like you working with a team but always talking with the team lead so you can think about it the same way. The squad by itself know how to operate inside and it has a surface to the Agent framework in. Net where the other parts of the application can talk with and everything just work together um, in producing better results over time.

Speaker A: Okay, so when Squad is uh, working on actually building software features that typically requires a very powerful model and uses a lot of tokens, but when it's just processing a paragraph and maybe doing a little bit of inference, it doesn't require as strong or as large of an LLM. And so what I'm thinking is if I'm going to deploy one of these applications using Squad, um, using the Agent framework. And now I have essentially I have a deployed application running if I deploy it to Azure. And now inside, uh, it's using Squad as part of the processing of the inference. It's potentially using a lot of tokens because now it's running a lot of transactions. Um, have you tested to see what the smallest model that still actually works is?

Speaker C: Yeah. So uh, two important things to note is that when we hire members to the squad, um, the Squad itself can also assign the relevant and the most appropriate model for each one of the sub agents and of course you can override it. So not all the agents will necessarily use the same model. Each one will pick and choose the right thing. Second, because we are running on top of Copilot and when I uh, explain about the agent, uh, framework we are running over the Copilot SDK which allow you to also bring your own P style of models. And these models can be of any type. I mean those can be really slm, like small models you run locally. They can be like heavy uh, LLMs running on foundry and you can just uh, try it out. So I did try with a few local models and got nice results. Nothing that like uh, it was publishing. Uh, we also had some community members that did something really awesome. You can define the budget for the different models, um, and the Squad itself will take that into consideration and have a fallback to lower tier uh, models when needed or based on the type of job it needs to do. And so it can adjust the models it's going to use based on the budget, the need, uh, and the entire orchestration. And so as we, you know, as the time goes and maybe by the time this episode is out, who knows what model will be uh, published and all the pricing will just change. I'm assuming for my own opinion that it's just a phase. Right now things looks really, really pricey. I'm sure that a year from now it's going to be a commodity. My mother will have the model running on our remote control for the tv. Uh, and maybe a Squad there as well. It's just a phase, uh, but I acknowledge it, uh, token, you need to have some kind of an economy. We try to give tools. You can also run in an economy model in Squad. Uh, but yeah, at the end it's kind of, you know, the lower thing you use like the performance and the quality of what you get will be affected. Uh, so there is a trade off. Um, I can't really recommend, um, but I'm guessing that over time we will find more ways to make it more balanced. Mhm.

Speaker A: Okay. And so with this episode, the Microsoft Build conference has happened just a couple of weeks ago. Um, at the time of this episode coming out with all those announcements and uh, there has been Foundry iq, Work IQ and all of the, all of the ways to create agents for completely non developers. This seems to be competing with those options which I think Microsoft has always had. There's always been, there's always been uh, like logic apps and uh, uh these ways to create something automatic and then you have the professional developer where you're actually going to put that code in Git source control. This track seems to be for the automation where the source code is going to be stored in a git source control as opposed to the Microsoft 365 where that code doesn't seem to be stored in a git repository just seems to be in Microsoft 365. Is that still the same where you have like the professional engineer track Source control driven. And then there's always going to be these other options that are not source control driven.

Speaker C: I'm not a prophet. We um, are now again in this type of a phase where we try to break all the boundaries. Um ah, Brady and I has this vision where we using SWOT as an engine for other products. Uh because we will see more and more of those use cases where you need teams of agents knowing how to work together. And so we are building uh in our eyes this type of a foundation for that. Um, we are communicating with a lot of different teams in Microsoft as well. Uh we can't know what will come. In a way it's like uh saying to a poet uh somebody already wrote a love song, don't write another love song. Right. So each one wants to write their version uh that they see how it fits their vision and how uh it plays with other stuff. Um and so I hope uh we'll be there uh for the long run. People like it, it's very popular. We have a very large community right now uh contributing and using that and we keep stretching it forward and we are not seeing ourselves as replacement for things m. Maybe something that knows how to work together. Uh so what I mentioned with the Agent framework allows you to, to have a single application where you have an agent which is squad and an agent which is maybe on uh Foundry an agent that it may be Something else. All of them working together. And so each one will. It's like the best of breed. Each one is good for the use cases and the type of work you want to do, and others are better for something else. At the end, I want to give the options, and, you know, people will choose what they like best. Mm.

Speaker A: Okay. Well, as a. As a final question, um, uh, if. If someone wants to get started with Squad right now, um, with. With the latest release that's out there, where should we point them?

Speaker C: So, uh, the first, uh, place they need to go is to our, uh, Squad website, which will share the link. Right. It's, uh, under Brady Gaster. Um, Brady is my, uh, pal, is my, uh, partner. For that, uh, is the entire reason, is the brain behind the Squad. Um, and together we do, uh, all the fun stuff. So Squad is inside the Brady gaster organization in GitHub. Uh, and from there you have all the docs, um, and the instruction to install the Squad cli. And then all you need to do is, uh, Squad, uh, init, inside your repo, and that's it. Talk with your squad, and every question that you have about Squad, just ask the squad. They will help you. They will, you know, guide you, and they do things for you. So that's part of the vision, right? Just use your squad to do the work that you don't want to do.

Speaker A: Awesome. Awesome. Well, Tamir, thanks so much for coming on the podcast and sharing, uh, with us everything going on with Squad agent workflows. I appreciate it.

Speaker C: Thank you. Thank you for having me.

Speaker A: My pleasure. My pleasure. And until next time, dear listener, keep shipping.

Speaker B: You've been listening to the AI DevOps podcast. You can find us on YouTube, Apple Podcasts, Google Play, and everywhere else. Visit our sponsor, ClearMeasure, at, uh, clearmeasure.com and on behalf of your host, Jeffrey Palermo, thanks for listening, and may God bless you.